Hey HN! My co-founder and I built Krnel-graph, a library for training and evaluating lightweight probes based on LLM representations. In our experience, we found that LLMs have enough general world knowledge that we can build lightweight signals on top of them, so we built this library to bring this concept to more developers.

This is a link to some tutorial content we wrote about how to use our library for LLM guardrailing. It turns out that LLM representations can make for more accurate guardrails than purpose-built guardrails like LlamaGuard, and this post shows you how!

We’re especially looking for feedback. This library is open-source: no ads, no logins, no telemetry (yet), all self-hosted for now.

My company, Krnel, is hoping to provide value-add services on top of an open core. Please reach out if you’re looking for advice or contracting engagements!